Reproduce “Adaptive-Policy Distillation: Knowledge Injection under Information Asymmetry” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hojun Cho, Kang Eun Jeon, Jaegul Choo
The paper introduces Adaptive-Policy Distillation (APD) for transferring document-specific facts into a model’s weights for later closed-book question answering. APD uses teacher–student divergence to adapt both the rollout mixture and the balance between reverse and forward KL loss. On QASPER, APD reaches 0.558 Answer-F1 versus 0.390 for on-policy distillation after 10 epochs, averaged over 20 documents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.